Need advice regarding anchor transformations for anchor-based object detection models

Hi there!

I am trying implement a anchor-based SSD (Single-Shot MultiBox Detector) within Sentis however I am a little confused about a few things to reduce latency as much as possible:

  • Is it recommended to do all the anchor based transformations using the Functional class within the Compile() method? Or would it be more efficient to get the outputs from the model and perform the conversion using a separate method?
  • I have tried implementing the Anchor-based approach using the Functional class, but I have failed to find a way to completely vectorize all the operations for faster processing.
  • Are there any pre-existing Anchor-related operations within the Functional class? Do you plan on adding anything regarding this in the future?
  • Could you recommend the most efficient way to define anchors and use them to transform the outputs from the model to actual bounding boxes (Generate Anchors before-hand, use specific methods etc.)?

Thanks in advance!